A chess experiment you can inspect.

What is LittlePly?

LittlePly is a small neural chess policy: it learns to predict moves from examples. It does not produce a reasoning trace. The playing system also uses chess rules and a separately bounded mate-search helper; a move from the bot is not always a move selected by the model.

Who is behind the project · The local research workstation

1. Play, or watch a recorded game

   Try CM-7 in a desktop WebGPU browser after a roughly 374 MB download, or challenge the Lichess bot. Check supported games and availability before challenging. The browser and Lichess use the same weights with different helper limits.

   Prefer to watch? Step through a CM-7 evaluation game. The archive includes losses and unfinished games, with checkpoint and helper settings.

2. Understand one move

   The learning guide explains tokens, the training objective and legal play. Then the architecture diagrams show the current system. CM-7 reads current board/state plus optional recent history. Models through CM-5 used full game history; these are different input families.

3. Read a result with its conditions

   The current evaluation records 21 wins, 19 draws, 58 losses and two unfinished games. CM-7 was adopted for its input structure and standalone-position training support, not a demonstrated strength win over CM-5. Different helper settings prevent a model-only comparison.

   Experiments are listed newest first. CM identifiers name checkpoints; EXP identifiers name training or evaluation experiments. Dates and hashes remain available as provenance.

4. Follow a claim into the implementation

   The source map connects data preparation, training, scoring and serving. Source pages have file navigation, search and stable line links. CM-7 reproduction distinguishes available weights, source and evidence from training payloads that are not bundled. Historical downloads remain versioned.

A few useful words

Checkpoint

Saved learned weights, identified by a CM number and exact hash.

Token / vocabulary / context

A token is one integer ID. The vocabulary counts available IDs; context counts ordered sequence positions. Neither is a count of chess moves. See the exact formats.

Ply

One player's move. A full chess move normally contains two plies.

Helper override

A verified mate-search move used instead of querying the model. See the decision branches.

Evaluation versus ongoing play

A frozen evaluation uses a recorded protocol. Continuously sampled local or Lichess games are exploratory observations, not the same benchmark. Use the archive's purpose, venue and engine filters.